Dynamic Content Popularity Prediction Based on Spatial-Temporal Feature in Vehicular Networks

Content popularity is crucial for edge caching in vehicular networks, where prediction accuracy directly impacts communication performance. To enhance prediction accuracy by fully exploiting the characteristics of content popularity, we propose a novel approach that integrates spatial correlation among roadside units and the temporal correlation of vehicle content requests. We jointly construct a spatial-temporal feature matrix using Euclidean distance and Spearman correlation, called the ES matrix, as the adjacency matrix of the graph structure. This ES matrix is integrated with Spatial-Temporal Transformer Network (STTN) to build the ES-STTNS model. To evaluate its effectiveness, we design a dynamic caching strategy that continuously updates RSU cache in real-time based on predicted content popularity, aiming to maximize Cache Hit Ratio (CHR). Extensive experiments demonstrate that ES-STTNS achieves highly accurate predictions for real vehicular content requests, reducing Mean Absolute Error (MAE) by up to 24.4% and Root Mean Square Error (RMSE) by up to 20.4% compared to temporal-only models, especially over longer prediction horizons, while achieving 90% CHR at higher cache capacities.